A New Approach to Face Recognition Based on Generalized Hough Transform and Local Image Descriptors

Date
2012-09
Authors
Moise, Marian
Journal Title
Journal ISSN
Volume Title
Publisher
Faculty of Graduate Studies and Research, University of Regina
Abstract

In this thesis a new approach to face recognition is presented. Face recognition is one of the most proli c research elds and also one of the most demanding. It is in uenced not only by face-related attributes such as pose, position, scale, facial expression, accessories and physiognomy changes, but also by environmental factors like illumination, background, occluding objects, and, lastly, camera characteristics. The generalized Hough transform is improved so that it can nd the image region that best matches the template face image. Its reference point and hit rate are later used to discriminate between faces. The transform takes into account not only the position of the points, but also the value of their corresponding descriptors, which are compared against one another using the matrix cosine similarity measure and contribute to the nal score. The proposed method does not require any training data and it can be extended to object recognition. Moreover, it embeds not only the local structure of the face, represented by the local descriptors, but also the global shape of the face, captured by the generalized Hough transform and used later to discriminate between faces. The main advantage of the new method stems from the fact that image descriptors better embed features than do simple pixels. As descriptors have higher memory requirements, they are computed only for the most interesting points in an image based on the Canny-edge detector. Based on the locally adaptive regression kernels descriptor, a new descriptor, namely, the gradient distance descriptor, is proposed in this work and test results for face identi cation using the Yale face database prove that it performs better than other descriptors. Moreover, the new method for face identi cation improves the recognition rate with at least 20% in comparison with Fisherfaces, no matter which descriptor is used. As there are a plenitude of descriptors described in the literature, the proposed method for face recognition can be further improved by combining multiple descriptors in order to provide better invariance to the a ne transformations and to increase the discriminative power.

Description
A Thesis Submitted to the Faculty of Graduate Studies and Research In Partial Fulfillment of the Requirements for the Degree of Master of Science in Computer Science, University of Regina. xi, 75 p.
Keywords
Citation
Collections